Ginther and Hayes 41
a small gender gap in average productivity consistent with that reported recently in The Chronicle of Higher Education
Schneider 1998. In order to evaluate changes in promotion by gender over time, longitudinal sample is divided into two cohorts
defined by the year an individual received their Ph.D.
III. Empirical Methodology
The study begins with an evaluation of the gender wage structure. Wage regressions are estimated as a function of demographic characteristics, aca-
demic background, employer characteristics, and academic productivity. The analy- sis continues by evaluating salary differentials using a salary decomposition devel-
oped by Oaxaca 1973 where the salary gap can be characterized as follows:
1 lnw
¯
m
⫺ lnw¯
f
⫽ ∆X¯′β
m
⫹ X¯ ′
f
∆β Let
∆X¯ ⫽ X¯
m
⫺ X¯
f
be the difference in average endowments and ∆β ⫽ β
m
⫺ β
f
be the differences in estimated coefficients salary structure, the term that accounts for the effect of discrimination.
In order to interpret coefficient differences as discrimination researchers must make several assumptions. First, one assumes that in the absence of discrimination,
the coefficients would be the same for men and women. Second, the model must contain all relevant explanatory variables. If some relevant variables are omitted,
then one cannot definitively argue that coefficient differences are due to discrimina- tion. Third, researchers much choose the nondiscriminatory salary structure. In Equa-
tion 1 we implicitly assume that the male coefficients represent the nondiscriminatory salary structure.
9
However, discrimination may operate by conferring unfair advan- tage in the form of nepotism to men and unfair disadvantage in the form of discrimi-
nation to women. Thus, the male salary structure may not represent the salary that would prevail in the absence of discrimination. Researchers including Neumark
1988 and Oaxaca and Ransom 1994 suggest using a weighted average of the male and female wage structure a pooled method to proxy for the unobserved
nondiscriminatory wage.
Although using the pooled method may correct for the combined effects of nepo- tism and discrimination, it requires additional assumptions about the unobserved
nondiscriminatory wage and the weighting mechanism used to obtain it. Even if the pooled approach provides an accurate estimate of the nondiscriminatory wage
structure, it cannot account for the competitive wage structure that would have pre- vailed had discrimination never existed Oaxaca and Ransom 1994. Finally, the
pooled approach is not likely to be used in legal cases concerned with discrimination. Oaxaca and Ransom observe 1994, p. 18: ‘‘In effect U.S. law leans toward the
adoption of the white or male wage structure as the norm.’’
Instead of using the pooled method to estimate the nondiscriminatory salary struc- ture, we use the male salary structure as the norm that would occur in the absence
of discrimination. Men are the comparison group used almost exclusively in legal proceedings of gender discrimination. In addition, using the male wage structure as
9. The researcher may also assume that the female coefficients represent the underlying salary structure.
42 The Journal of Human Resources
the norm does not require estimation of the unobserved counterfactual wage struc- ture. Thus, the coefficient estimates attributable to discrimination in this study may
reflect both male advantage and female disadvantage. As such, they may overstate the true effect of discrimination.
The study continues by evaluating gender differences in promotion using the lon- gitudinal sample and two empirical methods. First, we estimate probit models in
order to determine whether significant differences exist in the probability of promo- tion by gender. Second, duration models are used to estimate the conditional proba-
bility of promotion to tenure given the individual has survived untenured.
Duration to tenure is modeled using the proportional hazards model. The hazard function gives the instantaneous risk that promotion to tenure will occur at year t,
where the hazard of promotion h
i
t is a function of the baseline hazard λ
o
t and covariates, x in Equation 2.
2 h
i
t ⫽ λ
o
texp {β
1
x
i 1
⫹ ⋅ ⋅ ⋅ ⫹ β
k
x
ik
} The baseline hazard function is left unspecified and can be interpreted as the hazard
function for an individual whose covariates all equal zero. The covariates in Equation 2 influence the scale of the hazard rate and are not a function of time. Additional
covariates used in this analysis include demographic and employer characteristics, employment background, primary work activity, and productivity.
IV. Empirical Results
